A Late Answer Is a Wrong Answer: Heaps, Continuations and ETS on the BEAM

Blogs & Podcasts Blog Posts
, ,
1

Following up on Making `dns_erlang` fast: binary tricks for the BEAM , there’s our second post in our series on making erldns, an authoritative DNS server, faster: it’s techniques for hot paths that work in any Erlang or Elixir codebase, I really hope you’ll learn a trick or two.

  • Find, don’t count. ets:select_count/2 =/= 0 is the ETS version of length(L) =/= 0 when matching [_ | _] would do. Stopping at the first match cut our name servers’ CPU by a third.
  • One ets:next/2 instead of a match spec got many kinds of answers 1.8x faster. On an ordered_set, one step from a key that sorts just before the name tells you both whether the name exists and whether anything is below it.
  • Give short-lived processes the heap they need. Spawning a process with a well-chosen min_heap_size made a query take less than half the time as it ran to completion without a single GC cycle.

In production, p90 dropped by about 25%, and by about 33% if the query requests DNSSEC.

Questions and feedback welcome!

5 Likes